As companies move from single AI tools to multi-agent AI systems, monitoring becomes more complex. It is no longer sufficient to inspect the final AI output. Firms also need to understand which agent performed which task, what information each agent used, how agents communicated, where disagreements occurred, when humans intervened, and how errors propagated through the workflow.

This thesis investigates what organizations should monitor when AI work is performed by multiple interacting agents. Based on interviews with AI implementation leads, data scientists, compliance experts, or managers, the thesis develops a framework for monitoring multi-agent AI systems as organizational systems.

The thesis is supposed to answer the research question: "What should organizations monitor to govern multi-agent AI systems effectively?"

If you are interested in this thesis, feel free to contact Marc Grau (marcchristopher.grau@unisg.ch).